[Intel-gfx] [PATCH] Revert "drm/i915: Make prepare_plane_fb fully interruptible."

Ville Syrjälä ville.syrjala at linux.intel.com
Mon Nov 2 07:38:07 PST 2015


On Mon, Nov 02, 2015 at 01:41:03PM +0100, Maarten Lankhorst wrote:
> Hey,
> 
> Op 30-10-15 om 22:06 schreef ville.syrjala at linux.intel.com:
> > From: Ville Syrjälä <ville.syrjala at linux.intel.com>
> >
> > This reverts commit b26a6b35581c84124bd78b68cc02d171fbd572c9.
> >
> > commit b26a6b35581c ("drm/i915: Make prepare_plane_fb fully interruptible.")
> > breaks GPU reset on gen3/4 machines. Go back to to non-interruptible.
> >
> I've done some digging and by forcing an unconditional modeset during reset I was able to trigger it on my system.

Hmm, maybe we should add some kind of debug knob to do just that. That
way we could test most of the gen3/4 reset path with gen5+.

I thought we already had that for load detection, but I may have
imagined it considering that load detection seems to have been 
broken now for a while.

> It should be fixed by applying the rest of the interruptible
> series so we can mask EIO, followed by replacing i915_mutex_lock_interruptible with mutex_lock_interruptible so there will be no waiting for gpu reset.
> This is what's causing the deadlock here. :)
> 
> ~Maarten
> 
> diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
> index 22c0f8a54053..df6dbbc85855 100644
> --- a/drivers/gpu/drm/i915/intel_display.c
> +++ b/drivers/gpu/drm/i915/intel_display.c
> @@ -3207,9 +3207,11 @@ void intel_prepare_reset(struct drm_device *dev)
>  	if (IS_GEN2(dev))
>  		return;
>  
> +#if 0
>  	/* reset doesn't touch the display */
>  	if (INTEL_INFO(dev)->gen >= 5 || IS_G4X(dev))
>  		return;
> +#endif
>  
>  	drm_modeset_lock_all(dev);
>  	/*
> @@ -3245,7 +3247,12 @@ void intel_finish_reset(struct drm_device *dev)
>  		 * FIXME: Atomic will make this obsolete since we won't schedule
>  		 * CS-based flips (which might get lost in gpu resets) any more.
>  		 */
> +#if 0
>  		intel_update_primary_planes(dev);
> +#else
> +		intel_display_resume(dev);
> +		drm_modeset_unlock_all(dev);
> +#endif
>  		return;
>  	}
>  
> @@ -13174,12 +13181,12 @@ static int intel_atomic_prepare_commit(struct drm_device *dev,
>  			flush_workqueue(dev_priv->wq);
>  	}
>  
> -	ret = i915_mutex_lock_interruptible(dev);
> +	ret = mutex_lock_interruptible(&dev->struct_mutex);
>  	if (ret)
>  		return ret;
>  
>  	ret = drm_atomic_helper_prepare_planes(dev, state);
> -	if (!ret && !async) {
> +	if (!ret && !async && !i915_reset_in_progress(&dev_priv->gpu_error)) {
>  		u32 reset_counter;
>  
>  		reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);

-- 
Ville Syrjälä
Intel OTC


More information about the Intel-gfx mailing list